Model-Based Football Predictions Explained (xG, Elo, and What a Model Can’t Know)

xG and Elo answer different questions. Expected goals estimates the quality of recorded shots. Elo estimates a team’s relative strength from previous results and opponent quality. Neither metric independently predicts every event that can decide a football match.

Standard pre-shot xG assigns a scoring probability to each attempt by comparing it with similar historical shots. A close-range central attempt will normally receive a higher value than a difficult shot from distance or a narrow angle. Exact inputs and weightings vary between data providers.

A Simple xG Example

What xG Helps Reveal

• Chance quality: it separates a large number of weak shots from a smaller number of clear opportunities.
• Creation versus finishing: it helps distinguish whether a team produced good chances or converted difficult attempts.
• Defensive performance: xG conceded can show whether a defence repeatedly allowed high-quality opportunities.
• Scoreline context: it provides more information than goals alone when reviewing a match.

What xG Does Not Fully Capture

xG is not a complete measurement of attacking performance. A dangerous pass across goal that narrowly misses every attacker may create no recorded shot and therefore no standard xG value. Tactical control, pressing quality, positioning and game-state effects require separate analysis.

Elo is a continuously updated team-strength rating. A team gains rating points after performing better than expected and loses points after performing worse than expected. Defeating a highly rated opponent normally changes the rating more than beating a much weaker opponent.

Important Parts of a Football Elo Model

• Starting rating: the system needs an initial estimate before later results can update team strength.
• Opponent quality: results against stronger teams carry different information from results against weaker teams.
• Update speed: a K-factor or similar setting controls how quickly the rating reacts to each result.
• Home advantage: football models normally require a separate adjustment for the expected benefit of playing at home.
• Draw treatment: a full 1X2 forecast needs a method for estimating draw probability, not only a rating difference.

Where Elo Is Useful

Elo provides a relatively stable baseline when detailed shot or event data is unavailable. It can also reduce overreaction to one surprising score because every new result is interpreted in relation to the model’s previous expectation.

A football model estimates a distribution of possible outcomes. It does not identify one result that must happen. If a team is assigned a 60% win probability, the remaining 40% still represents draws and defeats.

Events That Can Change the Baseline

• Late lineup news: missing a goalkeeper, central defender, playmaker or leading scorer can materially change the match. Read more about how verified team news can change football probabilities .
• First goal: scoring first can alter possession, pressing behaviour, defensive risk and transition frequency.
• Red cards: an early dismissal can invalidate assumptions based on eleven-versus-eleven football.
• Set-piece swing: one corner or free kick can decide a low-scoring match even when open-play control favours the other team.
• Finishing and goalkeeping variance: a team can create better chances and still lose.

Accuracy alone can be misleading. A model that repeatedly chooses favourites may record many correct selections without producing well-calibrated probabilities. Strong evaluation considers both the outcome and the probability assigned before the match.

A model probability can be converted into model-implied decimal odds before bookmaker margin:

A 50% probability corresponds to decimal odds of 2.00. A 40% probability corresponds to 2.50. These figures describe the model’s estimate, not a guaranteed fair market price.

Why Bookmaker Margin Matters

The implied probabilities of all bookmaker prices in a market commonly add up to more than 100%. The amount above 100% represents the market margin. A direct comparison between one model probability and one displayed price should therefore account for the margin and for uncertainty in the model itself.

The video summarises the difference between probability and certainty, xG as a shot-quality measure and Elo as a long-term team-strength rating.

• Confirm whether the forecast uses current lineups or only historical team data.
• Check whether the model includes home advantage, draw probability and competition strength.
• Compare xG values from the same provider rather than combining incompatible sources.
• Review injuries, suspensions, rotation and schedule congestion before kickoff.
• Separate the probability of an outcome from whether the available price sufficiently compensates for risk.
• Avoid judging the model from one match; review calibration and performance across a large sample.

The explanations on this page are based on established football analytics and probability-forecasting concepts. Individual providers may use different datasets, variables and model specifications.

• Opta Analytics Football data, performance analysis and expected-goals applications.
• World Football Elo Ratings — About the Ratings An overview of Elo-style rating updates for international football.
• Scikit-learn — Brier Score Loss Documentation for evaluating the accuracy of probabilistic predictions.
• Scikit-learn — Log Loss Documentation for evaluating probability estimates and penalising confident errors.

xG is most useful for understanding the quality of recorded chances. Elo is most useful as a stable estimate of relative team strength. Combining the two can create a stronger baseline, but a complete football forecast also needs current lineups, home advantage, tactical context, draw modelling and uncertainty management.

The strongest model is not the one that claims certainty. It is the one that produces tested probabilities, explains its limitations and remains useful across a large sample of matches.